AI Glossary

An ordinary chatbot waits for each instruction. An agent can make a mini-plan, call a search tool, inspect the result, revise its approach, and continue until it reaches a stopping point.

Picture it
You say “compare these three contracts.” The agent opens the files, extracts clauses, builds a comparison, notices a missing page, and asks for it.

Related: tool use · inference · alignment

“Agentic” describes systems or tasks where the model repeatedly observes, decides, acts, and checks. It is a spectrum: a coding assistant that runs tests is mildly agentic; one that files a pull request after fixing failures is more so.

Shortcut
If the AI is doing a sequence of work rather than producing one response, someone will probably call it agentic.

Related: agent · tool use

Alignment includes following the user’s actual request, refusing harmful actions, staying honest about uncertainty, and not gaming the measurement used to evaluate it. The hard part is that human intent is often incomplete or contradictory.

Failure mode
“Make support calls shorter” should not cause an agent to hang up on customers. It met the metric but missed the intent.

Related: RLHF · hallucination

An AI API lets a developer send input—usually text, images, or audio—and receive a model’s output in a predictable format. It is the plumbing behind many AI features that do not look like chatbots.

Picture it
A help-desk app sends a ticket to a model API and gets back a draft reply plus a category label.

Related: inference · token

Benchmarks can test coding, math, factual recall, tool use, safety, or many other traits. A score is never the whole story: test design, prompting, contamination, and the software harness can all move the number.

Reading tip
“Best on Benchmark X” means “best under these test conditions,” not automatically “best for your job.”

In September 2026 reporting, DeepSeek’s own results changed materially with the harness; the figures remained vendor-reported. Source 1

Related: frontier model · hallucination

Asking a model to work through a problem can improve results, especially on multi-step tasks. But the displayed explanation is not a guaranteed window into the model’s true internal process; it can be incomplete or reconstructed.

Useful distinction
A convincing explanation and a correct answer are two different things. Check both.

Related: reasoning model · inference

Training is expensive and long, so teams save snapshots along the way. “Model checkpoint” is also commonly used to mean the downloadable model files you load to run or fine-tune a model.

Analogy
It is a save game for training—except the file may be hundreds of gigabytes.

Related: weights · fine-tuning

It includes your prompt, attached text, tool results, chat history, and often the model’s own output. A huge window means more fits, not that the model will use every part equally well.

September 2026 example
DeepSeek V4.1 Flash was reported with a one-million-token context window—an envelope large enough for a substantial codebase or many documents, though real retrieval quality still matters.

Source 2

Related: token · KV cache · RAG

Diffusion became a dominant approach for image generation and is also used for video, audio, and some language models. Generation starts from randomness, then progressively denoises toward something matching the prompt.

Analogy
Like seeing a photograph slowly emerge from television static—guided by a description.

Related: multimodal · foundation model

The student learns from outputs or signals produced by the larger model. It can become much cheaper and faster while retaining part of the teacher’s capability—but rarely all of it.

Analogy
Instead of reading the entire library, the student studies a carefully prepared set of worked examples.

Related: quantization · small language model

Similar items end up near one another in this mathematical space. Embeddings let software find conceptually related text even when it does not share the exact same words.

Picture it
A search for “late delivery” can retrieve a ticket saying “my package still hasn’t arrived.”

Related: RAG · token

Fine-tuning changes the model itself. It is useful for consistent style, format, vocabulary, or task behavior. It is not always the right way to add changing facts—that is often better handled with retrieval.

Picture it
A general model is fine-tuned on thousands of correctly labeled support tickets so it learns the company’s categories.

Related: RAG · weights · LoRA

Rather than building a separate model for summarizing, coding, classification, and chat, teams start with one large pretrained base and adapt or prompt it for each use.

Analogy
A versatile engine platform that can power a sedan, van, or race car after further engineering.

Related: fine-tuning · frontier model

“Frontier” is a moving, fuzzy category rather than a technical certification. It usually implies high capability, high training cost, and potentially greater safety scrutiny.

Reading tip
The frontier moves. A model described this way today may become a routine mid-tier option surprisingly quickly.

Related: foundation model · benchmark

GGUF bundles model tensors and metadata in a way tools such as llama.cpp can load. When you see several GGUF downloads labeled Q4, Q5, or Q8, those are usually different quantization levels.

In practice
You choose the GGUF variant that fits your memory budget, download it, and open it in a compatible local runner.

Related: quantization · local model

Language models generate likely continuations; they do not automatically check a fact database before every sentence. Retrieval, tools, citations, and verification can reduce errors, but none makes them impossible.

Red flag
A neat citation to a paper that does not exist is a classic hallucination. Confidence is style, not evidence.

Related: RAG · alignment · benchmark

Training teaches the model; inference is the work it does afterward. Each chatbot response, image generation, or speech transcription is an inference job. Cost and speed depend on model size, hardware, input length, and output length.

Analogy
Training is culinary school. Inference is cooking tonight’s dinner.

Related: token · KV cache · quantization

“KV” means key–value. The cache makes generation much faster, but it consumes memory and grows with context length. That is why very long conversations can strain a GPU even when the model itself fits.

Analogy
It is the model’s working scratchpad: faster than rereading every prior page, but the scratchpad takes desk space.

DeepSeek V4.1 Flash’s reported FP4 KV cache uses 890 bytes per token; that is a model-specific implementation detail, not a universal rate. Source 2

Related: context window · VRAM

An LLM learns statistical patterns from a vast amount of text and code. That next-token objective produces surprisingly general skills—but it does not make the model a database, a person, or automatically truthful.

Useful distinction
The chat app is the product interface. The LLM is one engine inside it.

Related: token · transformer · foundation model

Local inference can offer privacy, offline use, predictable cost, and low latency. The trade-offs are device limits, setup, battery or power use, and often lower capability than the largest hosted models.

September 2026 example
PrismML said its ternary Bonsai 2 compressed a 27B-parameter Qwen model to 5.9GB while retaining 98% of aggregate benchmark scores. Those are company-reported figures.

Source 3

Related: quantization · VRAM · small language model

Low-Rank Adaptation trains small add-on matrices while the base model stays frozen. The resulting adapter is much smaller to train, store, and swap than a full new checkpoint.

Analogy
A removable lens on a camera: it changes what the system is good at without rebuilding the camera.

Related: fine-tuning · weights

A router sends each token to a small set of “experts.” The model can have enormous total capacity without using every parameter on every step. It is still one trained system; the experts are not little human-like personalities.

September 2026 example
DeepSeek V4.1 Flash was reported as a 552B-parameter MoE with 8B active parameters per token—much less computation than activating all 552B at once.

Source 2

Related: parameter · inference

A multimodal model might accept text and images, or understand audio and video as well. “Multimodal” does not say how well it handles each mode, only that several are supported.

Picture it
You photograph a whiteboard, ask a spoken question, and receive a text summary.

Related: foundation model · token

Open weights are not automatically the same as open source. Training data, code, and recipes may remain closed; licenses may restrict commercial use. Always ask: what exactly is open, and under what terms?

September 2026 example
DeepSeek V4.1 Flash’s weights were released under the MIT license. Other “open” models may carry revenue thresholds or use restrictions.

Source 2

Related: weights · checkpoint · local model

Training adjusts billions of these numbers so the model produces useful outputs. Parameter count hints at capacity and hardware needs, but bigger does not automatically mean smarter: architecture, data, training, and evaluation matter.

Reading shorthand
“27B” means roughly 27 billion parameters. In an MoE model, “total” and “active per token” can be very different.

Related: mixture of experts · weights

A prompt may include user text, system rules, examples, retrieved documents, and tool output. Good prompting clarifies the task, context, constraints, and desired form—but cannot make a model know facts it has never seen.

Better than “summarize this”
“Give me the decision, three supporting facts, unresolved risks, and the owner of each next step.”

Related: context window · RAG

Think fewer digits after the decimal point. Lower-bit formats can shrink a model dramatically, sometimes with a modest quality cost. The best level depends on the model, task, hardware, and how much loss you can accept.

File-picker clue
Q8 is usually larger and closer to the original; Q4 is usually smaller and may lose more fidelity. “Usually” matters—formats and methods differ.

PrismML’s Bonsai 2 uses ternary weights (+1, −1, or 0), an especially aggressive form of low-precision representation. Its performance claims are company-reported. Source 3

Related: GGUF · VRAM · local model

RAG does not retrain the model. It adds selected source material to the prompt, often from search or a document database. Quality depends heavily on retrieving the right passages—not merely owning the right documents.

Picture it
Before answering an HR question, the system fetches the applicable handbook pages and asks the model to answer from them.

Related: embedding · hallucination · fine-tuning

Instead of answering immediately, it may generate and evaluate intermediate work. That can improve coding, math, and planning, while increasing latency and cost. It still makes mistakes and should not be treated as a formal proof system by default.

Trade-off
A hard debugging task may benefit from more “thinking”; rewriting a two-line email probably does not.

Related: chain of thought · inference

People compare outputs or provide ratings; those preferences help train a reward signal, which then shapes model behavior. Modern post-training uses several related techniques, so “RLHF” is sometimes used loosely for the whole alignment stage.

What it changes
Not the model’s raw ability to predict text so much as which answers it learns to prefer producing.

Related: alignment · fine-tuning

“Small” is relative and has no fixed cutoff. SLMs can be easier to run locally, cheaper to serve, and faster to specialize. For focused tasks, a good small model may beat a much larger general one on cost and latency.

Best fit
Classifying routine device events on a phone may not need a frontier model in a data center.

Related: local model · distillation

The model does not secretly gain direct access to a calendar, calculator, or browser. It produces a structured request; the surrounding application validates and executes it, then returns the result.

Picture it
The model asks for weather(city="Chicago"). The app calls the approved weather service and gives the response back to the model.

Related: agent · API · alignment

A token may be a whole short word, part of a longer word, punctuation, or whitespace. Models operate on tokens rather than words, and API pricing and context limits are usually measured in them.

Why it matters
Two passages with the same word count can use different numbers of tokens because languages, code, and formatting split differently.

Related: context window · inference

Its key mechanism, attention, lets the model weigh relationships among tokens. Transformers parallelize training effectively and scale well, which helped unlock today’s large models.

Don’t confuse it with
The sci-fi robot franchise. Here, “transformer” names a way of processing sequences.

Related: LLM · KV cache

The model’s weights, KV cache, and temporary working data need room. If they do not fit in VRAM, software may offload parts to system memory, which is typically slower.

Buying implication
For local AI, a GPU with more memory can matter more than a slightly faster GPU with less.

Related: quantization · KV cache · local model

Weights are parameters, viewed as the model artifact you save and load. The architecture is the blueprint; the weights are what training wrote into it. Possessing weights lets you run the model if you have compatible software and enough hardware.

Analogy
The architecture is an unprogrammed instrument. The weights are the tuning acquired through practice.

Related: parameter · open weights · checkpoint